Bug in training with small datasets and potentially with MIL

[dcifci]

While using for data size is equal to batch size, an empty dataloaders object is created without complaints, which leads a further error in the training phase.